Real Analysis Notes 2024/25 3.

1 Periodic functions

3 Fourier series
[References: Folland 8.1 - 2, Grafakos ’Classical Fourier Analysis’, 3.1, 3.2. (1 e 2) , 3.4 (1, 3, 4). Also
Katznelson, ’Introduction Harmonic Analysis’, I.1- 5, II.2. For pointwise convergence, ’Fourier analysis’, J.
Duoandikoetxea , Ch. 1. ]

The problem of representing a function f , typically defined on an interval of R, by a trigonometric

series of the form
1
X
f (x) = ak cos(kx) + bk sen(kx)
k=0

arises naturally when using the method of separation of variables to solve partial di↵erential equa-
tions. This appeared in the work of Fourier in the heat equation, and even earlier in work of Bernouilli
and Euler when studying the problem of a vibrating string.
Such a function f is necessarily periodic, as a sum of periodic functions with period 2⇡, and
therefore it suffices to consider f on any interval of length 2⇡. We consider here always intervals of
length 1, so we modify
n x 7! 2⇡x.
o Using Euler’s identity, e = cos(↵)+i sen(↵), we can replace sen(2⇡kx)
↵

and cos(2⇡kx) by e2⇡ikx : k 2 Z so that our problem is transformed to studying the representation of
f by
1
X
f (x) = ck e2⇡ikx .
k= 1

If x 2 Rn , we take k 2 Zn and k · x = k1 x1 + ... + kn xn , where k = (k1 , . . . , kn ). We also define

⇣ ⌘1/2
|k| = k12 + · · · + kn2 .

3.1 Periodic functions

We will consider functions on Rn that satisfy f (x + k) = f (x) for all x 2 Rn and k 2 Zn . Such functions
are called 1-periodic in every coordinate, so that it suffices to consider f on the cube [0, 1]n .
More precisely, if we identify points of the form x + k, then we can regard f as a function on a set
of equivalence classes: we say that x, y in Rn are equivalent and we write x ⌘ y if x y 2 Zn . It is
easy to check that ⌘ is an equivalence relation.
We define the n-torus Tn as the set Rn /Zn of all such equivalence classes, with the quotient
topology. This is nothing more than the cube [0, 1]n with opposite sides identified, that is, such
that points (x1 , . . . , xk , . . . , xn ) and (x1 , . . . , xk + 1, . . . , xn ) are identified. It is an additive group, with
identity element 0, which coincides with every e j = (0, . . . , 0, 1, 0, . . . , 0). We often think of Tn as the
set [ 1/2, 1/2]n .
If n = 1, then T can of course be regarded as the unit circle in C. In general, Tn can also be regarded
as
n⇣ ⌘ o
e2⇡ix1 , . . . , e2⇡ixn 2 Cn : (x1 , . . . , xn ) 2 [0, 1]n .

Functions that are 1-periodic on Rn can be identified with functions on Tn , so this will be our
domain for the remainder of this sections. It is a compact Hausdor↵ space.
The measure on Tn will be the restriction of n-dimensional Lebesgue measure to the set [0, 1]n ,
still denoted by dx. We will write
Z Z Z
f (x) dx = f (x)dx = f (x)dx
Tn [0,1]n [ 1/2,1/2]n

where on the integral on the left f is regarded as a function on Tn and on the integrals on the right, f
is regarded as a periodic function on Rn . We refer a few properties of integration in Tn that we will
use often:

69
Real Analysis Notes 2024/25 3.1 Periodic functions

1. Tn is a compact, Hausdor↵ space with m(Tn ) = 1. In particular, by Proposition 2.1.19,

we have nested inclusions

C(Tn ) ⇢ L1 (Tn ) ⇢ ... ⇢ Lp (Tn ) ⇢ ... ⇢ L1 (Tn ).

The inclusions Lq ⇢ Lp , 1  p < q  1, and C(Tn ) ⇢ Lp , 1  p < 1 are dense (Theorem

2.4.11).
2. Translation invariance of Lebesgue measure and the definition of functions on Tn give
that for all f 2 L1 (Tn ), we have
Z Z Z
f (x)dx = f (x)dx = f (x)dx
Tn [ 1/2,1/2]n [a1 ,1+a1 ]⇥···⇥[an ,1+an ]

for any real numbers a1 , . . . , an . In particular,

Z Z Z
⌧ y f (x)dx = f (x y)dx = f (x)dx, y 2 Tn .
Tn Tn Tn

3. Integration by parts: given f, g continuously di↵erentiable on Tn , the boundary terms

will cancel, hence we have
Z Z
@ j f (x)g(x)dx = @ j g(x) f (x)dx.
Tn Tn

4. Approximation of identity in Tn : (kr )r>0 such that

R
(i) Tn |kr (x)|dx < C, all r > 0,
R
(ii) Tn kr (x)dx = 1, all r > 0
R
(iii) <|x|<1/2 |kr (x)|dx ! 0, r ! 0.

Then Theorem 2.6.13 still holds. We will more often take a sequence (kN ) with N ! 1.

The functions ek (x) = e2⇡ik·x are obviously defined on Tn . Moreover they verify
|ek (x)| = 1, ek (x + y) = ek (x)ek (y), @ j ek (x) = (2⇡ik j )ek (x).

Definition 3.1.1. Let ck , k 2 Zn be a sequence of complex coefficients. A trigonometric polynomial

of degree N is a function of the form
X
PN (x) = ck e2⇡ik·x , x 2 Tn ,
P
|ki |N

where (ck ) is finitely supported. A trigonometric series is a series of the form

X
ck e2⇡ik·x .
k2Zn

Note that this is simply a formal definition, we are not assuming any type of convergence. The
next sections will be devoted to the issue of how can a function f be represented by such a series, that
is, in which way, and for which functions, can we write
X
f ⇠ ck e2⇡ik·x (7)
k2Zn

70
Real Analysis Notes 2024/25 3.2 Convergence in L2 -norm

and how do the coefficients ck relate to f .

Since any trigonometric polynomial is continuous, we have that if the series converges absolutely
for some x, then it converges uniformly, hence the limit function is necessarily continuous. This means
that if we are to consider more general classes of functions, we will be typically looking at conditional
convergence pointwise. There are a number of meanings that we can assign to the sum on the right,
and several types of convergence, depending not only on the class of functions f considered, but also
on how we decide to perform the infinite sum.

3.2 Convergence in L2 -norm

There is an important class of functions for which we have already studied this kind of question, as it
follows from the theory of Hilbert spaces. Consider the Hilbert space L2 (Tn ) and define ek (x) = e2⇡ik·x ,
k 2 Zn . Then we have
Z Z 8
>
2⇡i(k j)·x <1, k = j
>
hek , e j i = ek (x)e j (x) dx = e dx = >
>
Tn T n :0, k , j.

Hence, {ek }k2Zn is an orthonormal set. Note that the span of the set {ek }k2Zn coincides with the
trigonometric polynomials.

Theorem 3.2.1 (Density of trigonometric polynomials). The class of trigonometric polynomials is

dense in C(Tn ) with the uniform norm.

Proof. Exercise (for class). ⇤

Since we have a dense inclusion C(Tn ) ⇢ Lp (Tn ), 1  p < 1, it also follows:

Corollary 3.2.2. The class of trigonometric polynomials is dense in Lp (Tn ), 1  p < 1.

This means that any f 2 Lp (Tn ), and " > 0, we can find a trigonometric polynomial PN of degree
N large enough, such that
k f PN kp < ",
which is a step towards the existence of a representation as a trigonometric series. Moreover, for
p = 2, this result can be restated as saying that {ek }k2Zn is a complete set in the Hilbert space L2 (Tn ).
Since it is also orthonormal:

Corollary 3.2.3. The set {ek }k2Zn is a Hilbert basis for L2 (Tn ).

P
It follows that for any f 2 L2 (Tn ), the series k2Zn h f, ek iek converges to f in L2 (Tn ), independently
of the ordering of Zn , that is, f can be written as a trigonometric series
X X
f = h f, ek iek = fˆ(k)e2⇡ik·x (8)
k2Zn k2Zn

where fˆ(k) is the Fourier coefficient given by

Z
fˆ(k) = h f, ek i = f (x)e 2⇡ik·x
dx, (9)
Tn

and where the series on the right of (8) converges, in the L2 -norm, precisely when
X X
|h f, ek i|2 = | fˆ(k)|2 < 1
k2Zn k2Zn

71
Real Analysis Notes 2024/25 3.3 Fourier coefficients in L1

that is, if ( fˆ(k))k2Z 2 `2 (Zn ).

Theorem 3.2.4. The following hold for f, g 2 L2 (Tn ) :

1. The function f is a.e. equal to the L2 (Tn )-limit of the sequence
X
lim fb(k)e2⇡ikx .
N!1
|k|N

2. Plancherel’s identity: X
k f k22 = | fb(k)|2
k2Zn

3. Parseval’s identity: Z X
f (x)g(x)dx = fb(k)b
g(k)
Tn k2Zn

4. The map f 7! { fb(k)}k2Zn is an isometric isomorphism from L2 (Tn ) onto `2 (Zn ).

We call Fourier transform in L2 (Tn ) to the isomorphism

F : L2 (Tn ) ! `2 (Zn ), f 7! { fb(k)}k2Zn (10)

Since both Hilbert spaces are separable, the existence of such an isomorphism was established in
Section 1. The point here is that the Fourier transform provides an explicit description of such a map,
and that enables that we place questions in L2 (Tn ) in the space of frequencies `2 (Zn ).

3.3 Fourier coefficients in L1

In an e↵ort to extend the Fourier transform to a larger space, we first note that in fact the Fourier
coefficient in (9) is still well-defined, even if we consider f in L1 (Tn ).

Definition 3.3.1. For a function f in L1 (Tn ) and k in Zn , we define

Z
fb(k) = f (x)e 2⇡ik·x dx
Tn

We call fb(k) the k-th Fourier coefficient of f . The trigonometric series with these coefficients,
X
fˆ(k)e2⇡ikx ,
k2Zn

is called the Fourier series of f .

In T, we can check easily, using Euler’s identity, that such a series can be written in the form
1
A0 X
+ (Ak cos(2⇡kx) + Bn sen(2⇡kx))
2
k=1

where Ak = fˆ(k) + fˆ( k) and Bk = i( fˆ(k) fˆ( k)), corresponding to the form in which it appeared in
the original work of Fourier in the heat equation, and before that, in the wave equation. We think of
the numbers k as the frequencies.

72
Real Analysis Notes 2024/25 3.3 Fourier coefficients in L1

We see now some properties of Fourier coefficients. (The proofs are easy exercises.)

Proposition 3.3.2. Let f, g be in L1 (Tn ). Then for all m, k 2 Zn , we have

1. f[
+ g(k) = fb(k) + b
g(k); cf (k) = fb(k), 2 C;

2. If f¯(x) = f (x) denotes the complex conjugate of f , then b

f¯(k) = fb( k),
b
3. If fedenotes fe(x) = f ( x), then fe(k) = fb( k),

4. ⌧[ b
y ( f )(k) = f (k)e
2⇡ik·y
, y 2 Tn ,
⇣ ⌘^
5. e2⇡im(·) f (k) = fb(k m),
R
6. fb(0) = Tn
f (x)dx,

7. supk2Zn | fb(k)|  k f kL1 (Tn ) .

Remark 3.3.3. One can also define Fourier coefficients of a finite Borel measure µ on Tn : if k 2 Zn
then Z
µ(k) =
b e 2⇡ik·x dµ
Tn

is called the k th Fourier coefficient of µ.

Proposition 3.3.4. Let f, g be in L1 (Tn ). Then fd

⇤ g(k) = fb(k)b
g(k), k 2 Zn .

Proof. We have, applying Fubini’s theorem,

Z Z
[
f ⇤ g(k) = f (x y)g(y)e 2⇡ik·(x y)
e 2⇡ik·y
dydx = fb(k)b
g(k).
Tn Tn

⇤
The following property is the basis for applications of Fourier transform methods to di↵erential
equations.

Proposition 3.3.5. If f 2 C j (Tn ), for ↵ multi-index |↵|  j, any k 2 Zn ,

@d
↵ f (k) = (2⇡ik)↵ fb(k).

Proof. For j = 1, it follows by integration by parts. Then by induction. ⇤

We have then that the Fourier transform, defined as the map

F : L1 (Tn ) ! l1 (Zn ), f 7! fb

is a bounded linear operator with

kF ( f )kl1 (Zn )  k f kL1 (Tn ) ,
which restricts to an isomorphism L2 (Tn ) ⇢ L1 (Tn ) ! `2 (Zn ) ⇢ l1 (Zn ). In particular if f j ! f
in L1 , then fbj (k) ! fb(k) in l1 , that is, in the sup-norm. We have kF k = 1, since if f 0 then
ˆ
k f kL1 (Tn ) = f (0)  kF ( f )kl1 (Zn ) .

73
Real Analysis Notes 2024/25 3.4 Pointwise convergence

Moreover, F is an algebra morphism, between the Banach algebra L1 (Tn ) with convolution and
l (Zn ) with pointwise product, and it maps di↵erential operators in Ck (Tn ) to multiplication by
1

polynomials in l1 (Zn ).
We will see F is injective later, when we prove that we can always write f from its Fourier
coe↵cients. 10 .
Note that not every sequence in l1 (Zn ) is the Fourier transform of a function in L1 (Tn ), that is, the
Fourier transform is not surjective from L1 (T) to l1 (Z). In fact, we have the following fundamental
result (note that in L2 (Tn ) this follows from orthonormality).

Theorem 3.3.6 (Riemann-Lebesgue Lemma). If f 2 L1 (Tn ) and |k| = sup |k j |, a then

lim fb(k) = 0
|k|!1

a or any, necessarily equivalent, norm in Zn .

Proof. Use density of trigonometric polynomials in L1 (Tn ): given " > 0, take PN such that
c
k f PN k1 < ", then for |k| > N, have P N (k) = 0 hence

| fˆ(k)| = | fˆ(k) c
PN (k)|  k f PN k1 < ", for |k| > N.

⇤
Note that for n = 1, we can use Euler’s identity to get that, for f 2 L1 (T),
Z Z
lim f (t) sen(2⇡kt) dt = lim f (t) cos(2⇡kt) dt = 0.
|k|!1 T |k|!1 T

Remark 3.3.7 (Decay of Fourier coefficients). One important property of Fourier coefficients is that
their decay at infinity is linked to regularity: if f 2 C j (Tn ), then for |↵|  j, k , 0,

@d↵ f (k)
= fb(k),
(2⇡ik)↵

in particular, for f 2 C j (Tn ),

lim(1 + |k| j ) fb(k) ! 0, |k| ! 1.
See Grafakos 3.3 for further results. (Note that this yields that for j > 1, the Fourier series converges
absolutely.)

3.4 Pointwise convergence

We now want to study the convergence properties of Fourier series. For this, we shall now restrict to
n = 1 and consider the Fourier series of f 2 L1 (T)
1
X
fˆ(k)e2⇡ikx , k 2 Z, x 2 T
k= 1

We want to determine how and under which conditions this series represents the function f . We will
denote the N-th symmetric partial sum of the series by SN f (x);
N
X
SN f (x) = fˆ(k)e2⇡ikx .
k= N

10 Thereis an easy argument we can give now, using injectivity in L2 (Tn ): if fˆ(k) = ĝ(k), then ( f[g)(k) = 0 in l2 (Zn ), hence
f g = 0 in L2 (Tn ), that is, f = g a.e.

74
Real Analysis Notes 2024/25 3.4 Pointwise convergence

Then we have
N Z
X 1 Z 1 N
X
SN f (x) = f (t)e2⇡ik(x t) dt = f (t) e2⇡ik(x t) dt
k= N 0 0 k= N

Definition 3.4.1. The Dirichlet kernel is the trigonometric polynomial DN defined on T as

N
X
DN (t) = e2⇡ikt . (11)
k= N

Then we can write

Z 1 Z 1
SN f (x) = f (t)DN (x t)dt = f (x t)DN (t)dt = f ⇤ DN (x)
0 0

We have then that the convergence of SN f (x) relies on the convergence properties of DN . We have
Z 1
DN (t)dt = 1
0

ix ix
Noting that sen x = 2i(e e ), we find that
N
X
(e⇡it e ⇡it
) e2⇡ikt = e⇡i(2N+1)t e ⇡i(2N+1)t

k= N

to get

sen(⇡(2N + 1)t)
DN (t) = , t 2 T, t , 0.
sen(⇡t)

(or just do the sum of the geometric series). By definition, we have DN (0) = N + 1 = limt!0 DN (t),
hence DN is a continuous, even function. Moreover,
1
|DN (t)|  ,  |t|  1/2.
sen(⇡ )
However, Z 1/2
|DN (t)|dt C log N, as N ! 1.
1/2

(Exercise: compare with harmonic series.) In particular, DN is not an approximate identity. The
difficulty of studying pointwise convergence of Fourier series hinges on this fact. The numbers
LN := kDN kL1 , N = 1, 2, . . . are called the Lebesgue constants.
The following result can be regarded as a version of Riemann-Lebegue’s lemma.

1
Lemma 3.4.2. If f 2 L1 (T), then for any 2  a < b  12 , we have
Z b
lim f (t) sen(⇡(2N + 1)t)dt = 0.
N!1 a

1
Proof. It suffices to consider a = 2, b = 12 . Writing sen(⇡(2N + 1)t) = 1 ⇡(2N+1)it
2i (e e ⇡(2N+1)it
),
Z 1 Z 1
2 2
f (t) sen(⇡(2N + 1)t)dt = g1 (t)e2⇡Nit g2 (t)e 2⇡Nit
dt = gb1 ( N) gb2 (N)
1 1
2 2

75
Real Analysis Notes 2024/25 3.4 Pointwise convergence

where
1 1
f (t)e⇡it , g2 (t) = f (t)e ⇡it
g1 (t) =
2i 2i
are integrable. By the Riemann-Lebesgue lemma 3.3.6, the claim follows.
⇤
First we see that the convergence of Fourier series is in fact a local property: even if the Fourier
coefficients are computed from an integral in Tn , the convergence of the Fourier series at x relies only
on the behavior of f around x.This is really a consequence of the following:

Proposition 3.4.3. If f 2 L1 (T), then for any 0 <  12 , we have

Z
lim f (t)DN (t)dt = 0.
N!1 |t|>
Z
In particular, lim DN (t)dt = 0.
N!1 |t|>

Proof. Z Z 1/2
f (t)DN (t) = g(t) sen(⇡(2N + 1)t)dt
|t|> 1/2
where
f (t)
g(t) = { |t|<1/2} (t)
sen(⇡t)
is integrable, since | sen(⇡t)| > | sen(⇡ )|, for |t| > .
⇤

Theorem 3.4.4 (Riemann Localization Principle). If f 2 L1 (T) is such that f = 0 in a neighborhood

of x, then
lim SN f (x) = 0
N!1
1
In particular, if f, g 2 L (T) are such that f = g in a neighborhood of x, then their Fourier series at x
both converge or both diverge.

Proof. Suppose that f (t) = 0 on (x , x + ). Then

Z
lim SN f (x) = lim f (x t)DN (t)dt = 0.
N!1 N!1 |t|>

⇤
It follows from Proposition 3.4.3 that the convergence of SN f (x) relies on the behavior of the
integral Z Z
f (t)
f (t)DN (t)dt = sen(⇡(2N + 1)t)dt.
|t|< |t|< sen(⇡t)
We know that this integral goes to 0 as long as the function
f (t)
sen(⇡t)
is integrable in neighborhood of 0. Since sen(⇡t) ⇠ ⇡t as t ! 0, this issue can be reduced to
Lemma 3.4.5. For any integrable f 2 L1 (T), we have that the integrals
Z Z
f (t) f (t)
dt and dt
|t|< sen(⇡t) |t|< t
both diverge or both converge.

76
Real Analysis Notes 2024/25 3.4 Pointwise convergence

Proof. Exercise.

Theorem 3.4.6 (Dini’s Criterion). If for some x there exists > 0 such that
Z
f (x + t) f (x)
dt < 1
|t|< t

then
lim SN f (x) = f (x)
N!1

R
Proof. Since T
DN = 1 , and noting that DN is even, we can write
Z 1/2 Z 1/2
SN f (x) f (x) = f (x t) f (x) DN (t) dt = f (x + t) f (x) DN (t) dt
1/2 1/2
Z Z
f (x + t) f (x)
= sen(⇡(2N + 1)t) dt + f (x + t) f (x) DN (t) dt
|t|< sen(⇡t) |t|<1/2

By the results above, both of these integrals tend to 0 . ⇤

Remark 3.4.7. Note that we could have proved in the same way that if for some L
Z
f (x + t) L
dt < 1
|t|< t

then
lim SN f (x) = L.
N!1

Dini’s criterion applies In particular to Lipschitz functions, or more generally, if f is Hölder

continuous then its Fourier series converges pointwise to f . (For continuous functions this result is
not true, as we will note below.) Recall that f 2 C0,↵ (T) is Hölder- ↵ continuous if

| f (x) f (y)|  C|x y|↵ ,

for some C 0 and all x, y 2 T. In this case,

f (x + t) f (x) C
 ,
t |t|1 ↵
which is integrable in |t| < , for any 0 < ↵  1, independently of x. This yields that the convergence
below is in fact uniform (Exercise).

Corollary 3.4.8. Let f 2 C0,↵ (T) be Hölder- ↵ continuous, for 0 < ↵  1. Then the Fourier series of f
converges to f on T, and
N
X
f (x) = lim SN ( f )(x) = lim fˆ(k)e2⇡ikx .
N!1 N!1
N

In Dini’s criterion above, we could have replaced f (x) by any L 2 C. Using the fact that DN is
even, we can write Z 1/2
SN f (x) = ( f (x + t) + f (x t))DN (t)dt,
0

which yields an alternative formulation of Dini’s criterion given above, also commonly known as
Dini’s criterion or Dini - Dirichlet criterion, and that can be proved exactly in the same way.

77
Real Analysis Notes 2024/25 3.4 Pointwise convergence

Theorem 3.4.9 (Dini - Dirichlet Criterion). If for some x there exists > 0 such that
Z
f (x + t) + f (x t) 1
L dt < 1
0 2 t

then limN!1 SN f (x) = L.

Probably the first known result on the pointwise convergence of Fourier series was given by
Dirichlet. A function f is said to be piecewise C1 in T if there exist side derivatives at each point, in
particular, also exist side limits f (x± ) = limh!0+ f (x ± h), and

f (x ± h) f (x± )
lim+
h!0 h
exist in C.

Theorem 3.4.10 (Dirichlet’s Criterion). If f is piecewise C1 in T, then its Fourier series converges
to 12 ( f (x+ ) + f (x )),
1
lim SN f (x) = ( f (x+ ) + f (x )).
N!1 2
In particular, if f is continuous at x, then SN f (x) ! f (x).

Dirrichlet’s original result was in fact for functions that are piecewise monotonic. This result was
extended later by Jordan to functions with bounded variation. Recall that such a function f can be
written as the di↵erence of two monotonic functions; in particular f (x+ ) and f (x ) exist everywhere.

Continuous functions
Let now f 2 C(T). If f verifies any of the conditions given in the previous section, so that Dini’s
criterion applies, then its Fourier series converges to f (x). We also know from L2 -convergence that if
f 2 C(T) ⇢ L2 (T), there is a subsequence of the partial sums SN j ( f ) that converges to f a.e.
However, given a continuous function f , we cannot in general guarantee that its Fourier series
converges pointwise.

Theorem 3.4.11 (du Bois-Reymond). There exists (a dense subset of) continuous functions whose
Fourier series diverges at a point (in fact, at a dense set).

For a constructive example see eg Grafakos - Ex.3.4.7. The existence of continuous functions
whose Fourier series diverges at a given point x can be shown using the divergence of the L1 -norm
of the Dirichlet kernel, together with an important result in Functional Analysis, called the Uniform
Boundedness principle, also known as the Banach-Steinhaus theorem (Folland Thm. 5.13, Rudin Thm.
5.8): given Tn : X ! Y a sequence of bounded operators from a Banach space X to a normed space Y,
then if (Tn f )n2N is bounded for all f 2 X then the operator norm kTn k is bounded.

R A common argument goes as follows: let x = 0 and consider TN f := SN ( f )(0) = f ⇤ DN (0) =

T
f (t)DN (t)dt. Then TN is a family of bounded functionals on the complete space C(T), with

kTN k = kDN k1 = LN .

Since kTN k ! 1, by the Uniform boundedness principle, there must be f 2 C(T) such that |TN f | ! 1
hence SN f does not converge at 0.
In fact, it can be proved that there is a dense set of such functions f where SN f (0) does not converge
(see [Rudin Real and complex Analysis], 5.11 - 5.14), and we can argue in the same way for any x 2 T.

78
Real Analysis Notes 2024/25 3.5 Summability methods

Convergence in Lp -norm
We have seen that due to results from Hilbert spaces, the Fourier series of an f 2 L2 (T) always
converges to f in L2 - norm. The issue is whether this result also holds in Lp if p , 2.
We have that SN f converges to f in Lp norm for all f 2 Lp (T), 1  p < 1, if and only if there exists
Cp such that
SN f p
 Cp k f kp , 8N, f 2 Lp (T).

The necessity also follows from the Uniform Boundedness principle. We leave it as an easy
exercise to show it is sufficient (use density of trigonometric polynomials).
If 1 < p < 1, then this inequality holds: this is a consequence of an important result in Harmonic
Analysis on the boundedness of the conjugate operator (see Grafakos 4.1 or Katznelson).
When p = 1, the L1 operator norm of SN coincides with kDN k1 = LN ! 1, and so we cannot
guarantee convergence. In fact, Kolmogorov gave an example of an integrable function whose
Fourier series diverges at every point.
As for almost everywhere convergence, a deep result by Carleson (1965) showed that if f 2 L2 (T),
then its Fourier series indeed converges almost everywhere. This was extended to f 2 Lp , 1 < p < 1
by Hunt.

3.5 Summability methods

In order to recover a function f from its Fourier coefficients, it would be convenient to find other
method than taking the limit of the partial sums of its Fourier series. A useful point in this respect is
that the convolution of functions f 2 L1 (T) with any trigonometric polynomial is again a trigonometric
polynomial of degree N.

Proposition 3.5.1. If f 2 L1 (Tn ) and PN is a trigonometric polynomial of degree N with coefficients

ck then f ⇤ PN is a trigonometric polynomial of degree N with coefficients

⇤ PN (k) = ck fˆ(k).
f[

R
Proof. Just note that, writing ek (x) = e2⇡ikx then f ⇤ ek (x) = Tn
f (t)e2⇡ik(x t) dt = fˆ(k)e2⇡ikx . ⇤

So we can think of replacing SN f with f ⇤KN , with KN an approximate identity made of trigonomet-
ric polynomials. Such functions KN are called summability kernels. Note that we will be considering
series of the form
X1
ck,N fˆ(k)e2⇡ikt
1

and the point is to take ck,N that ensure better convergence properties, and such that limN!1 ck,N = 1.
One way to do this is to consider the averages of SN . Recall that if a sequence (ak ) converges to
a, then its arithmetic means a1 +...+a
k
k
also converges to a, the converse being not true (as ak = ( 1)k
shows). So, instead of considering the partial sum SN f we can consider instead the Cesaro means of f
given by
1
N ( f )(t) = S0 f (t) + ... + SN f (t) . (12)
N+1
We have then that
N ( f )(t) !` ) SN f (t) ! `.
Since SN f = f ⇤ DN , for any N 2 Z, we have that

1
N ( f )(t) = f ⇤ FN (t), where FN (t) := (D0 (t) + ... + DN (t)) .
N+1

79
Real Analysis Notes 2024/25 3.5 Summability methods

Definition 3.5.2. The Fejér kernel is the function FN defined as on T

N
1 X
FN (t) = Dk (t)
N+1
k=0

We have ! !2
N
X |k| 2⇡ikt 1 sen(⇡(N + 1)t)
FN (t) = 1 e = .
N+1 N+1 sen(⇡t)
k= N

The first equality follows from rewriting the sum

N
X N X
X k N
X
Dk (t) = e2⇡ikt = (N + 1 |k|) e2⇡ikt ,
k=0 j=0 k= k k= N

To check the last equality for x , 0 we use that

⇣ ⌘
sen2 (x) = 2 e2ix e 2ix
/4

and check that

⇣ N
⌘X
2 e2⇡it e 2⇡it
N+1 | j| e2⇡i jt = 2 e2⇡i(N+1)t e 2⇡i(N+1)t
.
j= N

It follows that limt!0 FN (t) = N+1 and FN is continuous at 0, since it is the sum of continuous functions,
with FN (0) = N + 1. It is also an even function, with FN 0, and a trigonometric polynomial of degree
N with coefficients !
c |k|
FN (k) = 1 , |k|  N.
N+1
Hence N( f ) = f ⇤ FN is also a trigonometric polynomial
N
X !
|k|
N( f ) = fb(k) 1 e2⇡ikt .
N+1
k= N

The crucial point is that the Fejér kernel is an approximate identity: it is straightforward that
Z
FN (t) dt = 1
T
R
(since DN = 1) and since FN 0, also kFN k1 = 1. Moreover, for each 0 <  1/2,
Z
2
FN (t)dt  ! 0, N!1
1/2>|t|> (N + 1) sen2 (⇡ )

(note sen(⇡t) sen(⇡ ), < t < 1/2). We then have the following fundamental result on the
convergence of the Cesaro means:

Theorem 3.5.3 (Fejér’s theorem). The sequence {FN }N2N is an approximate identity as N ! 1 and
therefore the Cesaro means N ( f ) satisfy:
1. if f 2 Lp (T), then N( f ) = f ⇤ FN (t) ! f in Lp (T), 1  p < 1;

2. if f 2 C(T), then N( f ) = f ⇤ FN (t) ! f uniformly in C(T).

80
Real Analysis Notes 2024/25 3.5 Summability methods

A corollary of Féjer’s theorem is an alternative proof that the trigonometric polynomials are dense
in C(T) and in Lp (T). Moreover, recalling that if SN ( f ) converges also its average N ( f ) converges, we
have:

Corollary 3.5.4. Let f 2 L1 (T). If SN ( f ) converges in Lp (T) or in C(T), then its limit is f .

A consequence of Fejér’s theorem is that the sequence of Fourier coefficients determines f uniquely
a.e.

Proposition 3.5.5. If f, g 2 L1 (T) are such that fb(k) = b

g(k) for all k in Z, then f = g a.e.

Proof. By linearity, it suffices to assume that g = 0. If fb(k) = 0 for all k 2 Z, then

N
X !
|k|
N ( f ) = f ⇤ FN (t) = fb(k) 1 e2⇡ikt = 0
N+1
k= N

for all N 2 Z+ . Hence, since

f FN ⇤ f L1
!0
as N ! 1, so k f kL1 = 0, from which we conclude that f = 0 a.e. ⇤
In particular, we obtain a (partial) inversion results:

Proposition 3.5.6 (Fourier inversion). Suppose that f 2 L1 (T) and that

X
| fb(k)| < 1.
k2Z

Then X
f (x) = fb(k)e2⇡ikx a.e.,
k2Z

and therefore f is a.e. equal to a continuous function.

Proof. Both functions are well defined and have the same Fourier coefficients, hence they are a.e
equal. Moreover, the function on the right is everywhere continuous. ⇤
The class A(T) := { f 2 L1 (T) : fˆ 2 l1 (Z)} is called the Wiener algebra. It is closed for pointwise
products and convolution product, and it can be seen that

C1 (T) ⇢ A(T) ⇢ C(T)

1
and that if f is Hölder continuous with exponent 2 < ↵  1 then f 2 A(T). (See Katznelson 1.6,
Grafakos 3.3.3).
As for pointwise convergence of Cesaro means, we have now the following result, which weakens
the condition given by Dini’s criterion,

1
Theorem
⇣ ⌘ ⇣ ⌘(Fejér). If a function f 2 L (T) has left and right limits at a point x0 , denoted by
3.5.7
f x0 and f x0 , respectively, then
+

1
N ( f )(x0 ) = FN ⇤ f (x0 ) ! f (x0 +) + f (x0 ) as N!1
2
In particular, this holds for functions of bounded variation.

81
Real Analysis Notes 2024/25 3.5 Summability methods

Proof. Exercise (for class.)

Corollary 3.5.8. If f 2 C(T) is such that the partial sums SN f (x) converge for some x 2 T, then
limN!1 SN ( f )(x) = f (x).

82